| import torch | |
| class SimAgentActor: | |
| """Base class for GPUDrive torch simulation agents. | |
| Args: | |
| is_controlled_func (torch.Tensor): Determines which agents are controlled by this actor (across worlds). | |
| valid_agent_mask (torch.Tensor): Mask that determines which agents are valid, and thus controllable, in the environment. Shape: (num_worlds, num_agents). | |
| device (str): The device. | |
| """ | |
| def __init__(self, is_controlled_func, valid_agent_mask, device="cuda"): | |
| self.is_controlled_func = is_controlled_func | |
| self.device = device | |
| self.valid_and_controlled_mask = self.get_valid_actor_mask( | |
| is_controlled_func, valid_agent_mask | |
| ) | |
| self.actor_ids = [ | |
| torch.where(self.valid_and_controlled_mask[world_idx, :])[0] | |
| for world_idx in range(valid_agent_mask.shape[0]) | |
| ] | |
| def select_action(self, obs) -> torch.Tensor: | |
| """Select an action based on an observation. | |
| Args: | |
| obs (torch.Tensor): Batch of observations of shape (num_samples, observation_dim). | |
| Returns: | |
| torch.Tensor: _description_ | |
| """ | |
| raise NotImplementedError | |
| def get_valid_actor_mask(self, is_controlled_func, valid_agent_mask): | |
| """Returns a boolean mask across worlds that indicates which agents | |
| are valid and controlled by this actor. | |
| """ | |
| num_worlds = valid_agent_mask.shape[0] | |
| is_controlled_func = is_controlled_func.expand((num_worlds, -1)) | |
| assert ( | |
| is_controlled_func.shape == valid_agent_mask.shape | |
| ), f"is_controlled_func and valid_agent_mask must match but are not: {is_controlled_func.shape} vs {valid_agent_mask.shape}" | |
| return is_controlled_func.to(self.device) & valid_agent_mask.to( | |
| self.device | |
| ) | |